#74
Phi-3.5, L4 on GCloud, ADAS, GPT-4o Finetuning, OP Fraud-proof reverts, Network School, McD shitcoins, Appchains, L1 execution, handwriting, vertical-align, GDocs future bug, svgl, pglite & more
👋🏻 Welcome to the 74th!
📰 Read #74 on Substack for the best formatting
What’s happening 📰
☁️ ICYMI, In a leaked recording, AWS’s CEO tells employees that most developers could stop coding as soon as AI takes over. However, they are not the first ones to tell us this, being a software engineer in 2025, will be different than being one in 2020. Also, nothing to panic about, at the end of the day it’s about obsessions (read off-topic section for this).
✨ AGI Digest
⚓️ Model and Leaderboard Drops
🦍 The Gorilla-LLM team at UC Berkley released a new version of their function calling leaderboard, the BFCL V2 • Live which contains live1, user-contributed function documentation, and queries, avoiding the drawbacks of dataset contamination and biased benchmarks. What does it mean for you? Since there is less contamination and the dataset is being constantly updated, this represents real-life scenarios much better than the previous version. Though of course, the best evals for you are the ones that you make specific to your own needs!
The GPT-4 family reigns at function calling in overall accuracy! ɸ Microsoft silently released the Phi-3.5 family of SLMs consisting of Phi-3.5-mini (3.8B), Phi-3.5-MoE (16 3.8B experts, a total of 42B parameters and 6.6B active with two experts), and Phi-3.5-vision (4.2B). While all of these models punch above their weights, the MoE is particularly important because of the speed-quality improvements it brings because of the virtue of its architecture. If you’re using Gemini-1.5-Flash for your vision tasks, you might want to drop in the vision model and see how it fares in your tasks.
🎉 AI21 Labs released the Jamba 1.5 family of models — the Jamba 1.5 Mini (12B active/52B total) and the Jamba 1.5 Large (94B active/398B total) which are based on the SSM-Transformer architecture which weaves together Transformer’s quality with Mamba’s efficiency resulting in improvement in long context handling, speed, and quality. This is the first time a non-transformer model has performed so competitively on the lmsys (and other) benchmarks.
🦙 Google Cloud allows attaching Nvidia L4 GPUs to its serverless cloud offering Google Cloud Run. This means you can quickly deploy small models such as the Gemma 2B or LLama-3-8B via llama.cpp or ollama in your cloud without worrying about the infrastructure and getting started quickly.
✍️ Researchers from the Vector Institute released ADAS - Automated Design of Agentic Systems that aims to automatically design powerful agents & a new method where a meta agent invents new agents by programming them in code. This allows it to create and/or use any conceivable component that iteratively programs new agents based on a growing archive of previous designs.
🔧 OpenAI announced GPT-4o Finetuning with 1M training tokens per day for free for every organization through September 23. psstt… it was actually a finetuned GPT-4o model used by Genie to get to the top of the SWE-Bench leaderboard (we covered Genie last week).
🔐 0x Digest
⏮️ Just two months after Optimism’s permissionless fraud proofs went live on June 10, the foundation has reverted the network to a permissioned state after "community-driven audits" identified several bugs at varying levels of severity
🏫 Balajis is making his “Network State” dream come true by announcing The Network School, a new school near Singapore for the dark talent of the world. The 3-month-long pop-up city with four goals: learn, burn, earn, and have fun.
💰 This week in “Never a dull day in Crypto”, we have McDonald’s Instagram page hacked by crypto scammers who shilled a Soalan shitcoin & claims that they stole $700K.
🐣 Some “interesting takes” from the Bird app this week
David Hoffman from Bankless raised a point that “we should reprioritize L1 execution efforts” and it’s been ~2 years that we have prioritized L2 execution, though this has its importance, but has led to stagnancy on L1 execution improvements.
Pavel (not Durov) wrote, “Is everything really moving towards AppChains?” and the answer is “Well, it depends” (and other variants of it).
Movement Labs posted an explainer video where they mistakenly used “instant finality” instead of “fast” and “getting security from their own validator set”, and ended up getting trashed on are they a Rollup?
🪙 a16z Crypto wrote a “Your guide to tokens: How to design, launch, structure rights, and more”. Eventually either you ship a product or a token and rarely both. In all cases, it’s a good post to go through, it’s a compilation of their Token Launch Playbook.
🛠️ Dev & Design Digest
✍🏻 A new web standard mechanism is proposed to indicate whether an element or its subtree should allow user agent handwriting input. HTML Attribute to allow or disallow handwriting on demand, just like
contenteditablefor the editors. [Explainer here]🧮 CSS finally added vertical centering in 2024 in the default layout (flow). Before the introduction of
align-content, developers had to use workarounds like creating tables or using flexbox or grid layouts to align content vertically.⛓️💥 GitHub’s markdown renderer does not include the target attribute even if you use plain HTML. So you can’t open a link in a new tab in a README.
⚙️ Bun tweeted about their "drop-in" replacement for the "setup-node" GitHub action step with "setup-bun". They move fast.
What brings us to awe 😳
🌞 Reflect Orbital, a company that sells sunlight using a constellation of in-space reflectors. SaaS (Sunlight as a service) was unexpected.
🤯 In TypeScript, You can pass a string template literal into Pick/Omit. Though we have used both features separately, we never considered using them together.
😭 In a funny series of events, Node.js is taking steps towards removing Corepack from its distribution, aiming for changes in the next major release. If you remember, late in 2023, someone raised a PR to enable corepack by default, funnily enough, that issue brought eyes to the use of corepack and led to this day.
Source: Twitter 🤡 The “The 1 Hour per Year Bug” in Google Docs from back in 2021 and how daylight savings, ironically doesn’t save any developer hours. The bug was due to daylight savings time ending and users creating comments around that time being shown that it was created “tomorrow”. (comments in the future)
Today I (we) Learnt 📑
☕️ Some grammar lessons, we stumbled upon What's wrong with constructions like "Dragons are big, green, and eat people."?
This pattern is called a series out of control or bastard enumeration. It belongs to the broader class of failures of parallel structure such as "Mary likes hiking, swimming, and riding a bicycle.", where the first two items of the list are gerunds2 but the last is an infinitive.
💥 Saying “smash that like button” in a YouTube video makes it glow for the viewer!
🎵 The national anthems of Spain, Bosnia and Herzegovina, Kosovo, and San Marino have no lyrics. (just music, how the anthems should have been if Uncle Charlie made them) [Source: Chitthi]
🤝 You have read ~50% of Nibble, the following section brings tools out from the wild.
What we have been trying 🔖
📚 Abbey: An all-in-one platform that organizes learning materials, simplifies quiz creation, and supports course development.
📑 anthropics/courses: Education Prompt Engineering resources by Anthropic.
🖼️ svgl: Crowd-sourced repository of useful SVGs. We recently came to know there is a Raycast extension for the same.
📒 srcbook: TypeScript notebook for rapid prototyping, like Jupytr Notebook but for JS/TS.
Builders’ Nest 🛠️
💽 pglite: Lightweight Postgres packaged as WASM into a TypeScript library for the browser, Node.js, Bun, and Deno.
🗝️ local-vault: Store key-value data encrypted (biometric passkey protected), locally in the client.
😼 kitty: Cross-platform, fast, feature-rich, GPU-based terminal.
⛏️ llm-scraper: Turn any webpage into structured data using LLMs
Meme of the week 😌
Off-topic reads/watches 🧗
⛵️ The prevailing conditions by Seth Godin, a lot of conditions might feel like wind (out of control), but knowing it’s possible to change them makes all the difference.
⏰ Software estimates have never worked and never will by DHH. Well, I know a few people who are estimating using better methods and mathematics and it’s kinda working for them.
💪🏻 Obsession by Phil Eaton, where he shares the importance of obsession in achieving big goals and making long-term commitments and why they are important.
😶🌫️ You Are NOT Dumb, You Just Lack the Prerequisites by @lelouchdaily is a nice read about why people feel dumb when doing something complex and realize they are not good enough. (*ahem* imposter syndrome *ahem*).
Wisdom Bits 👀
"Progress is being aware when there is a storm happening inside of you and remaining calm as it passes by."
― Yung Pueblo
Wallpaper of the week 🌁
🌌 Grab the week’s wallpaper at wow.nibbles.dev
Weekly Standup 🫠
Nibbler A had a fun week starting with Rakshabandhan, serious building mid-week, and ending the week watching Inside Out 2 with homies. He says he did play some 🏸 this week. (maybe?)
Nibbler P had a lot of sweets this Rakshabandhan and made a small escapade out of his little town during the rains to watch Alien: Romulus.
If you liked what you just read, recommend us to a friend who’d love this too 👇🏻
“live” here means constantly updating
which is to be carried out